load model from drive and convert

.gitignore

@@ -0,0 +1 @@
+checkpoint-*/

README.md

@@ -0,0 +1,55 @@
+---
+license: other
+tags:
+- generated_from_trainer
+model-index:
+- name: opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8
+
+This model is a fine-tuned version of [facebook/opt-2.7b](https://huggingface.co/facebook/opt-2.7b) on the None dataset.
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 4e-05
+- train_batch_size: 8
+- eval_batch_size: 8
+- seed: 42
+- distributed_type: multi-GPU
+- gradient_accumulation_steps: 16
+- total_train_batch_size: 128
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_ratio: 0.01
+- num_epochs: 3
+
+### Training results
+
+
+
+### Framework versions
+
+- Transformers 4.19.2
+- Pytorch 1.10.0+cu113
+- Datasets 2.2.2
+- Tokenizers 0.12.1

config.json

@@ -0,0 +1,28 @@
+{
+  "_name_or_path": "facebook/opt-2.7b",
+  "activation_dropout": 0.0,
+  "activation_function": "relu",
+  "architectures": [
+    "OPTForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "do_layer_norm_before": true,
+  "dropout": 0.1,
+  "eos_token_id": 2,
+  "ffn_dim": 10240,
+  "hidden_size": 2560,
+  "init_std": 0.02,
+  "layerdrop": 0.0,
+  "max_position_embeddings": 2048,
+  "model_type": "opt",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "pad_token_id": 1,
+  "prefix": "</s>",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.19.2",
+  "use_cache": false,
+  "vocab_size": 50265,
+  "word_embed_proj_dim": 2560
+}

latest

@@ -0,0 +1 @@
+global_step1852

opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8_training_metadata.json

@@ -0,0 +1 @@
+{"output_dir": "/content/drive/MyDrive/Programming/hf-trainer/opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8", "overwrite_output_dir": true, "do_train": false, "do_eval": false, "do_predict": false, "evaluation_strategy": "no", "prediction_loss_only": false, "per_device_train_batch_size": 8, "per_device_eval_batch_size": 8, "per_gpu_train_batch_size": "None", "per_gpu_eval_batch_size": "None", "gradient_accumulation_steps": 16, "eval_accumulation_steps": 8, "eval_delay": 0, "learning_rate": 4e-05, "weight_decay": 0.1, "adam_beta1": 0.9, "adam_beta2": 0.999, "adam_epsilon": 1e-08, "max_grad_norm": 1, "num_train_epochs": 3, "max_steps": -1, "lr_scheduler_type": "cosine", "warmup_ratio": 0.01, "warmup_steps": 0, "log_level": -1, "log_level_replica": -1, "log_on_each_node": true, "logging_dir": "/content/drive/MyDrive/Programming/hf-trainer/opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8/logs", "logging_strategy": "steps", "logging_first_step": false, "logging_steps": 5, "logging_nan_inf_filter": true, "save_strategy": "epoch", "save_steps": 500, "save_total_limit": 1, "save_on_each_node": false, "no_cuda": false, "seed": 42, "data_seed": "None", "bf16": true, "fp16": false, "fp16_opt_level": "O1", "half_precision_backend": "amp", "bf16_full_eval": true, "fp16_full_eval": false, "tf32": "None", "local_rank": 0, "xpu_backend": "None", "tpu_num_cores": "None", "tpu_metrics_debug": false, "debug": "[]", "dataloader_drop_last": false, "eval_steps": "None", "dataloader_num_workers": 0, "past_index": -1, "run_name": "/content/drive/MyDrive/Programming/hf-trainer/opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8", "disable_tqdm": false, "remove_unused_columns": true, "label_names": "None", "load_best_model_at_end": false, "metric_for_best_model": "None", "greater_is_better": "None", "ignore_data_skip": false, "sharded_ddp": "[]", "fsdp": "[]", "fsdp_min_num_params": 0, "deepspeed": "ds_config_zero2_bf16.json", "label_smoothing_factor": 0.0, "optim": "adamw_hf", "adafactor": false, "group_by_length": false, "length_column_name": "length", "report_to": "['tensorboard']", "ddp_find_unused_parameters": "None", "ddp_bucket_cap_mb": "None", "dataloader_pin_memory": true, "skip_memory_metrics": true, "use_legacy_prediction_loop": false, "push_to_hub": true, "resume_from_checkpoint": "None", "hub_model_id": "opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8", "hub_strategy": "end", "hub_token": "<HUB_TOKEN>", "hub_private_repo": false, "gradient_checkpointing": true, "include_inputs_for_metrics": false, "fp16_backend": "auto", "push_to_hub_model_id": "None", "push_to_hub_organization": "None", "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>", "_n_gpu": 1, "mp_parameters": "", "auto_find_batch_size": false, "full_determinism": false, "train_batch_size": 8, "eval_batch_size": 8, "configs_src": "opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8", "data_tag": "text-file-input"}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69aff7264edf04861c8f3b6eed3f553e343af5ee8aebeb5ad1635f39ad2b4683
+size 10606359699

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "eos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "unk_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "pad_token": "</s>"}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "bos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "eos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "add_prefix_space": false, "errors": "replace", "pad_token": {"content": "<pad>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "add_bos_token": true, "special_tokens_map_file": null, "name_or_path": "facebook/opt-2.7b", "model_max_length": 512, "tokenizer_class": "GPT2Tokenizer"}

trainer_state.json

@@ -0,0 +1,2245 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 2.9985838559579205,
+  "global_step": 1851,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "learning_rate": 1.0526315789473684e-05,
+      "loss": 11.5,
+      "step": 5
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 2.105263157894737e-05,
+      "loss": 6.7393,
+      "step": 10
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.157894736842106e-05,
+      "loss": 3.7191,
+      "step": 15
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.999997059313686e-05,
+      "loss": 3.1877,
+      "step": 20
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 3.999894136200706e-05,
+      "loss": 2.9055,
+      "step": 25
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 3.9996441874195635e-05,
+      "loss": 2.9141,
+      "step": 30
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 3.999247231345674e-05,
+      "loss": 2.7834,
+      "step": 35
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 3.998703297161948e-05,
+      "loss": 2.7901,
+      "step": 40
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 3.9980124248566466e-05,
+      "loss": 2.8676,
+      "step": 45
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 3.9971746652204386e-05,
+      "loss": 2.6787,
+      "step": 50
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 3.996190079842669e-05,
+      "loss": 2.704,
+      "step": 55
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 3.995058741106831e-05,
+      "loss": 2.6286,
+      "step": 60
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 3.993780732185244e-05,
+      "loss": 2.6983,
+      "step": 65
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 3.992356147032939e-05,
+      "loss": 2.6522,
+      "step": 70
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 3.9907850903807514e-05,
+      "loss": 2.6293,
+      "step": 75
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 3.989067677727622e-05,
+      "loss": 2.5399,
+      "step": 80
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 3.987204035332105e-05,
+      "loss": 2.6687,
+      "step": 85
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 3.985194300203087e-05,
+      "loss": 2.6291,
+      "step": 90
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 3.983038620089714e-05,
+      "loss": 2.6533,
+      "step": 95
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 3.980737153470528e-05,
+      "loss": 2.5923,
+      "step": 100
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 3.97829006954182e-05,
+      "loss": 2.5941,
+      "step": 105
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 3.9756975482051855e-05,
+      "loss": 2.6612,
+      "step": 110
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 3.972959780054306e-05,
+      "loss": 2.6058,
+      "step": 115
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 3.9700769663609304e-05,
+      "loss": 2.5226,
+      "step": 120
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 3.967049319060081e-05,
+      "loss": 2.5573,
+      "step": 125
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 3.963877060734473e-05,
+      "loss": 2.603,
+      "step": 130
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 3.9605604245981515e-05,
+      "loss": 2.6506,
+      "step": 135
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 3.9570996544793445e-05,
+      "loss": 2.631,
+      "step": 140
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 3.9534950048025396e-05,
+      "loss": 2.5668,
+      "step": 145
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 3.9497467405697756e-05,
+      "loss": 2.6354,
+      "step": 150
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 3.9458551373411664e-05,
+      "loss": 2.5427,
+      "step": 155
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 3.941820481214637e-05,
+      "loss": 2.5586,
+      "step": 160
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 3.937643068804896e-05,
+      "loss": 2.5577,
+      "step": 165
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 3.933323207221624e-05,
+      "loss": 2.5664,
+      "step": 170
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 3.9288612140468984e-05,
+      "loss": 2.5396,
+      "step": 175
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 3.924257417311846e-05,
+      "loss": 2.5558,
+      "step": 180
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 3.919512155472529e-05,
+      "loss": 2.5306,
+      "step": 185
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 3.9146257773850585e-05,
+      "loss": 2.4945,
+      "step": 190
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 3.9095986422799506e-05,
+      "loss": 2.6086,
+      "step": 195
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 3.904431119735718e-05,
+      "loss": 2.4973,
+      "step": 200
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 3.899123589651695e-05,
+      "loss": 2.5872,
+      "step": 205
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 3.893676442220114e-05,
+      "loss": 2.5216,
+      "step": 210
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 3.888090077897418e-05,
+      "loss": 2.5367,
+      "step": 215
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 3.882364907374819e-05,
+      "loss": 2.5495,
+      "step": 220
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 3.8765013515481065e-05,
+      "loss": 2.6037,
+      "step": 225
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 3.870499841486705e-05,
+      "loss": 2.5012,
+      "step": 230
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 3.864360818401982e-05,
+      "loss": 2.4773,
+      "step": 235
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 3.8580847336148105e-05,
+      "loss": 2.52,
+      "step": 240
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 3.851672048522395e-05,
+      "loss": 2.4718,
+      "step": 245
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 3.845123234564343e-05,
+      "loss": 2.5251,
+      "step": 250
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 3.838438773188014e-05,
+      "loss": 2.5547,
+      "step": 255
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 3.831619155813119e-05,
+      "loss": 2.475,
+      "step": 260
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 3.8246648837955965e-05,
+      "loss": 2.4957,
+      "step": 265
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 3.817576468390753e-05,
+      "loss": 2.5425,
+      "step": 270
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 3.810354430715678e-05,
+      "loss": 2.5495,
+      "step": 275
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 3.802999301710932e-05,
+      "loss": 2.5375,
+      "step": 280
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 3.795511622101516e-05,
+      "loss": 2.5151,
+      "step": 285
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 3.787891942357115e-05,
+      "loss": 2.4864,
+      "step": 290
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 3.780140822651633e-05,
+      "loss": 2.5396,
+      "step": 295
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 3.77225883282201e-05,
+      "loss": 2.5585,
+      "step": 300
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 3.764246552326328e-05,
+      "loss": 2.4947,
+      "step": 305
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 3.756104570201213e-05,
+      "loss": 2.5004,
+      "step": 310
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 3.747833485018529e-05,
+      "loss": 2.4729,
+      "step": 315
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 3.739433904841375e-05,
+      "loss": 2.4764,
+      "step": 320
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 3.7309064471793794e-05,
+      "loss": 2.5369,
+      "step": 325
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 3.7222517389433085e-05,
+      "loss": 2.4836,
+      "step": 330
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 3.7134704163989705e-05,
+      "loss": 2.4558,
+      "step": 335
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 3.7045631251204434e-05,
+      "loss": 2.5562,
+      "step": 340
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 3.6955305199426164e-05,
+      "loss": 2.4603,
+      "step": 345
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 3.6863732649130426e-05,
+      "loss": 2.4176,
+      "step": 350
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 3.677092033243128e-05,
+      "loss": 2.4991,
+      "step": 355
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 3.667687507258631e-05,
+      "loss": 2.4846,
+      "step": 360
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 3.658160378349508e-05,
+      "loss": 2.5998,
+      "step": 365
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 3.648511346919079e-05,
+      "loss": 2.5061,
+      "step": 370
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 3.638741122332539e-05,
+      "loss": 2.4538,
+      "step": 375
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 3.628850422864807e-05,
+      "loss": 2.498,
+      "step": 380
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 3.618839975647718e-05,
+      "loss": 2.4867,
+      "step": 385
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 3.608710516616575e-05,
+      "loss": 2.5426,
+      "step": 390
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 3.598462790456035e-05,
+      "loss": 2.4842,
+      "step": 395
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 3.588097550545368e-05,
+      "loss": 2.4274,
+      "step": 400
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 3.5776155589030725e-05,
+      "loss": 2.5354,
+      "step": 405
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 3.5670175861308496e-05,
+      "loss": 2.455,
+      "step": 410
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 3.556304411356954e-05,
+      "loss": 2.5039,
+      "step": 415
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 3.545476822178915e-05,
+      "loss": 2.518,
+      "step": 420
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 3.5345356146056326e-05,
+      "loss": 2.4718,
+      "step": 425
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 3.52348159299886e-05,
+      "loss": 2.4318,
+      "step": 430
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 3.512315570014071e-05,
+      "loss": 2.5146,
+      "step": 435
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 3.50103836654071e-05,
+      "loss": 2.4801,
+      "step": 440
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 3.489650811641849e-05,
+      "loss": 2.4429,
+      "step": 445
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 3.478153742493235e-05,
+      "loss": 2.4654,
+      "step": 450
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 3.4665480043217444e-05,
+      "loss": 2.4846,
+      "step": 455
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 3.454834450343245e-05,
+      "loss": 2.4371,
+      "step": 460
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 3.443013941699868e-05,
+      "loss": 2.4071,
+      "step": 465
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 3.431087347396702e-05,
+      "loss": 2.3886,
+      "step": 470
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 3.419055544237906e-05,
+      "loss": 2.4539,
+      "step": 475
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 3.40691941676225e-05,
+      "loss": 2.4406,
+      "step": 480
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 3.394679857178086e-05,
+      "loss": 2.3687,
+      "step": 485
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 3.382337765297756e-05,
+      "loss": 2.4244,
+      "step": 490
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 3.3698940484714394e-05,
+      "loss": 2.4477,
+      "step": 495
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 3.35734962152045e-05,
+      "loss": 2.5432,
+      "step": 500
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 3.34470540666998e-05,
+      "loss": 2.4633,
+      "step": 505
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 3.331962333481302e-05,
+      "loss": 2.4302,
+      "step": 510
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 3.319121338783428e-05,
+      "loss": 2.4528,
+      "step": 515
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 3.3061833666042416e-05,
+      "loss": 2.3741,
+      "step": 520
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 3.29314936810109e-05,
+      "loss": 2.4638,
+      "step": 525
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 3.280020301490863e-05,
+      "loss": 2.4204,
+      "step": 530
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 3.2667971319795473e-05,
+      "loss": 2.436,
+      "step": 535
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 3.253480831691264e-05,
+      "loss": 2.4194,
+      "step": 540
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 3.240072379596806e-05,
+      "loss": 2.3565,
+      "step": 545
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 3.226572761441666e-05,
+      "loss": 2.4421,
+      "step": 550
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 3.2129829696735636e-05,
+      "loss": 2.4169,
+      "step": 555
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 3.1993040033694916e-05,
+      "loss": 2.4425,
+      "step": 560
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 3.1855368681622584e-05,
+      "loss": 2.399,
+      "step": 565
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 3.171682576166565e-05,
+      "loss": 2.3747,
+      "step": 570
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 3.1577421459045905e-05,
+      "loss": 2.437,
+      "step": 575
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 3.143716602231122e-05,
+      "loss": 2.4131,
+      "step": 580
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 3.129606976258201e-05,
+      "loss": 2.4329,
+      "step": 585
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 3.115414305279327e-05,
+      "loss": 2.4521,
+      "step": 590
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 3.101139632693197e-05,
+      "loss": 2.3317,
+      "step": 595
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 3.086784007926996e-05,
+      "loss": 2.4119,
+      "step": 600
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 3.072348486359247e-05,
+      "loss": 2.4315,
+      "step": 605
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 3.0578341292422286e-05,
+      "loss": 2.4687,
+      "step": 610
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 3.043242003623947e-05,
+      "loss": 2.4703,
+      "step": 615
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 3.0285731822696954e-05,
+      "loss": 2.5997,
+      "step": 620
+    },
+    {
+      "epoch": 1.01,
+      "learning_rate": 3.0138287435831855e-05,
+      "loss": 2.0103,
+      "step": 625
+    },
+    {
+      "epoch": 1.02,
+      "learning_rate": 2.9990097715272694e-05,
+      "loss": 1.9907,
+      "step": 630
+    },
+    {
+      "epoch": 1.03,
+      "learning_rate": 2.9841173555442463e-05,
+      "loss": 1.9292,
+      "step": 635
+    },
+    {
+      "epoch": 1.04,
+      "learning_rate": 2.9691525904757745e-05,
+      "loss": 1.9898,
+      "step": 640
+    },
+    {
+      "epoch": 1.05,
+      "learning_rate": 2.954116576482378e-05,
+      "loss": 1.9234,
+      "step": 645
+    },
+    {
+      "epoch": 1.05,
+      "learning_rate": 2.9390104189625702e-05,
+      "loss": 1.8726,
+      "step": 650
+    },
+    {
+      "epoch": 1.06,
+      "learning_rate": 2.923835228471587e-05,
+      "loss": 1.9208,
+      "step": 655
+    },
+    {
+      "epoch": 1.07,
+      "learning_rate": 2.90859212063974e-05,
+      "loss": 1.9407,
+      "step": 660
+    },
+    {
+      "epoch": 1.08,
+      "learning_rate": 2.8932822160904038e-05,
+      "loss": 1.9377,
+      "step": 665
+    },
+    {
+      "epoch": 1.09,
+      "learning_rate": 2.877906640357628e-05,
+      "loss": 1.9665,
+      "step": 670
+    },
+    {
+      "epoch": 1.09,
+      "learning_rate": 2.862466523803393e-05,
+      "loss": 1.9723,
+      "step": 675
+    },
+    {
+      "epoch": 1.1,
+      "learning_rate": 2.846963001534507e-05,
+      "loss": 1.9876,
+      "step": 680
+    },
+    {
+      "epoch": 1.11,
+      "learning_rate": 2.8313972133191615e-05,
+      "loss": 1.9405,
+      "step": 685
+    },
+    {
+      "epoch": 1.12,
+      "learning_rate": 2.8157703035031353e-05,
+      "loss": 1.9848,
+      "step": 690
+    },
+    {
+      "epoch": 1.13,
+      "learning_rate": 2.8000834209256665e-05,
+      "loss": 1.9328,
+      "step": 695
+    },
+    {
+      "epoch": 1.13,
+      "learning_rate": 2.7843377188349962e-05,
+      "loss": 1.9343,
+      "step": 700
+    },
+    {
+      "epoch": 1.14,
+      "learning_rate": 2.768534354803581e-05,
+      "loss": 1.9245,
+      "step": 705
+    },
+    {
+      "epoch": 1.15,
+      "learning_rate": 2.752674490642996e-05,
+      "loss": 1.9526,
+      "step": 710
+    },
+    {
+      "epoch": 1.16,
+      "learning_rate": 2.7367592923185207e-05,
+      "loss": 1.93,
+      "step": 715
+    },
+    {
+      "epoch": 1.17,
+      "learning_rate": 2.720789929863421e-05,
+      "loss": 1.9263,
+      "step": 720
+    },
+    {
+      "epoch": 1.17,
+      "learning_rate": 2.7047675772929328e-05,
+      "loss": 1.9432,
+      "step": 725
+    },
+    {
+      "epoch": 1.18,
+      "learning_rate": 2.6886934125179504e-05,
+      "loss": 1.9481,
+      "step": 730
+    },
+    {
+      "epoch": 1.19,
+      "learning_rate": 2.672568617258432e-05,
+      "loss": 1.909,
+      "step": 735
+    },
+    {
+      "epoch": 1.2,
+      "learning_rate": 2.6563943769565258e-05,
+      "loss": 1.9386,
+      "step": 740
+    },
+    {
+      "epoch": 1.21,
+      "learning_rate": 2.6401718806894144e-05,
+      "loss": 1.9362,
+      "step": 745
+    },
+    {
+      "epoch": 1.22,
+      "learning_rate": 2.6239023210819027e-05,
+      "loss": 1.9494,
+      "step": 750
+    },
+    {
+      "epoch": 1.22,
+      "learning_rate": 2.6075868942187366e-05,
+      "loss": 1.9576,
+      "step": 755
+    },
+    {
+      "epoch": 1.23,
+      "learning_rate": 2.5912267995566746e-05,
+      "loss": 1.937,
+      "step": 760
+    },
+    {
+      "epoch": 1.24,
+      "learning_rate": 2.5748232398363044e-05,
+      "loss": 1.9889,
+      "step": 765
+    },
+    {
+      "epoch": 1.25,
+      "learning_rate": 2.5583774209936218e-05,
+      "loss": 1.9285,
+      "step": 770
+    },
+    {
+      "epoch": 1.26,
+      "learning_rate": 2.5418905520713767e-05,
+      "loss": 1.895,
+      "step": 775
+    },
+    {
+      "epoch": 1.26,
+      "learning_rate": 2.525363845130185e-05,
+      "loss": 1.9826,
+      "step": 780
+    },
+    {
+      "epoch": 1.27,
+      "learning_rate": 2.5087985151594235e-05,
+      "loss": 1.9869,
+      "step": 785
+    },
+    {
+      "epoch": 1.28,
+      "learning_rate": 2.4921957799879076e-05,
+      "loss": 1.9325,
+      "step": 790
+    },
+    {
+      "epoch": 1.29,
+      "learning_rate": 2.4755568601943615e-05,
+      "loss": 1.9479,
+      "step": 795
+    },
+    {
+      "epoch": 1.3,
+      "learning_rate": 2.4588829790176837e-05,
+      "loss": 1.9616,
+      "step": 800
+    },
+    {
+      "epoch": 1.3,
+      "learning_rate": 2.4421753622670178e-05,
+      "loss": 1.9706,
+      "step": 805
+    },
+    {
+      "epoch": 1.31,
+      "learning_rate": 2.425435238231638e-05,
+      "loss": 1.9675,
+      "step": 810
+    },
+    {
+      "epoch": 1.32,
+      "learning_rate": 2.4086638375906484e-05,
+      "loss": 1.9684,
+      "step": 815
+    },
+    {
+      "epoch": 1.33,
+      "learning_rate": 2.3918623933225043e-05,
+      "loss": 1.9388,
+      "step": 820
+    },
+    {
+      "epoch": 1.34,
+      "learning_rate": 2.375032140614372e-05,
+      "loss": 1.9326,
+      "step": 825
+    },
+    {
+      "epoch": 1.34,
+      "learning_rate": 2.3581743167713187e-05,
+      "loss": 1.9521,
+      "step": 830
+    },
+    {
+      "epoch": 1.35,
+      "learning_rate": 2.3412901611253524e-05,
+      "loss": 1.9704,
+      "step": 835
+    },
+    {
+      "epoch": 1.36,
+      "learning_rate": 2.3243809149443077e-05,
+      "loss": 1.89,
+      "step": 840
+    },
+    {
+      "epoch": 1.37,
+      "learning_rate": 2.3074478213405937e-05,
+      "loss": 1.9438,
+      "step": 845
+    },
+    {
+      "epoch": 1.38,
+      "learning_rate": 2.2904921251798052e-05,
+      "loss": 1.9682,
+      "step": 850
+    },
+    {
+      "epoch": 1.39,
+      "learning_rate": 2.2735150729892013e-05,
+      "loss": 2.008,
+      "step": 855
+    },
+    {
+      "epoch": 1.39,
+      "learning_rate": 2.2565179128660667e-05,
+      "loss": 1.9247,
+      "step": 860
+    },
+    {
+      "epoch": 1.4,
+      "learning_rate": 2.2395018943859558e-05,
+      "loss": 1.9377,
+      "step": 865
+    },
+    {
+      "epoch": 1.41,
+      "learning_rate": 2.222468268510828e-05,
+      "loss": 1.9396,
+      "step": 870
+    },
+    {
+      "epoch": 1.42,
+      "learning_rate": 2.2054182874970808e-05,
+      "loss": 1.9848,
+      "step": 875
+    },
+    {
+      "epoch": 1.43,
+      "learning_rate": 2.188353204803486e-05,
+      "loss": 1.9382,
+      "step": 880
+    },
+    {
+      "epoch": 1.43,
+      "learning_rate": 2.1712742749990444e-05,
+      "loss": 1.9431,
+      "step": 885
+    },
+    {
+      "epoch": 1.44,
+      "learning_rate": 2.154182753670749e-05,
+      "loss": 1.9833,
+      "step": 890
+    },
+    {
+      "epoch": 1.45,
+      "learning_rate": 2.1370798973312813e-05,
+      "loss": 1.9338,
+      "step": 895
+    },
+    {
+      "epoch": 1.46,
+      "learning_rate": 2.1199669633266353e-05,
+      "loss": 1.9543,
+      "step": 900
+    },
+    {
+      "epoch": 1.47,
+      "learning_rate": 2.102845209743682e-05,
+      "loss": 1.9455,
+      "step": 905
+    },
+    {
+      "epoch": 1.47,
+      "learning_rate": 2.085715895317679e-05,
+      "loss": 1.9533,
+      "step": 910
+    },
+    {
+      "epoch": 1.48,
+      "learning_rate": 2.0685802793397317e-05,
+      "loss": 2.0128,
+      "step": 915
+    },
+    {
+      "epoch": 1.49,
+      "learning_rate": 2.051439621564216e-05,
+      "loss": 1.9471,
+      "step": 920
+    },
+    {
+      "epoch": 1.5,
+      "learning_rate": 2.0342951821161648e-05,
+      "loss": 1.9474,
+      "step": 925
+    },
+    {
+      "epoch": 1.51,
+      "learning_rate": 2.017148221398625e-05,
+      "loss": 1.9946,
+      "step": 930
+    },
+    {
+      "epoch": 1.51,
+      "learning_rate": 2e-05,
+      "loss": 1.913,
+      "step": 935
+    },
+    {
+      "epoch": 1.52,
+      "learning_rate": 1.9828517786013752e-05,
+      "loss": 1.981,
+      "step": 940
+    },
+    {
+      "epoch": 1.53,
+      "learning_rate": 1.965704817883836e-05,
+      "loss": 1.9809,
+      "step": 945
+    },
+    {
+      "epoch": 1.54,
+      "learning_rate": 1.948560378435784e-05,
+      "loss": 1.9793,
+      "step": 950
+    },
+    {
+      "epoch": 1.55,
+      "learning_rate": 1.9314197206602693e-05,
+      "loss": 1.9207,
+      "step": 955
+    },
+    {
+      "epoch": 1.56,
+      "learning_rate": 1.914284104682322e-05,
+      "loss": 1.8926,
+      "step": 960
+    },
+    {
+      "epoch": 1.56,
+      "learning_rate": 1.897154790256319e-05,
+      "loss": 2.0005,
+      "step": 965
+    },
+    {
+      "epoch": 1.57,
+      "learning_rate": 1.8800330366733654e-05,
+      "loss": 1.9432,
+      "step": 970
+    },
+    {
+      "epoch": 1.58,
+      "learning_rate": 1.862920102668719e-05,
+      "loss": 1.8667,
+      "step": 975
+    },
+    {
+      "epoch": 1.59,
+      "learning_rate": 1.8458172463292516e-05,
+      "loss": 1.9405,
+      "step": 980
+    },
+    {
+      "epoch": 1.6,
+      "learning_rate": 1.828725725000956e-05,
+      "loss": 1.9617,
+      "step": 985
+    },
+    {
+      "epoch": 1.6,
+      "learning_rate": 1.8116467951965145e-05,
+      "loss": 1.9447,
+      "step": 990
+    },
+    {
+      "epoch": 1.61,
+      "learning_rate": 1.79458171250292e-05,
+      "loss": 1.9093,
+      "step": 995
+    },
+    {
+      "epoch": 1.62,
+      "learning_rate": 1.7775317314891724e-05,
+      "loss": 1.9051,
+      "step": 1000
+    },
+    {
+      "epoch": 1.63,
+      "learning_rate": 1.7604981056140446e-05,
+      "loss": 1.916,
+      "step": 1005
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 1.7434820871339336e-05,
+      "loss": 1.8569,
+      "step": 1010
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 1.7264849270107994e-05,
+      "loss": 1.9163,
+      "step": 1015
+    },
+    {
+      "epoch": 1.65,
+      "learning_rate": 1.709507874820195e-05,
+      "loss": 1.9342,
+      "step": 1020
+    },
+    {
+      "epoch": 1.66,
+      "learning_rate": 1.6925521786594067e-05,
+      "loss": 1.8947,
+      "step": 1025
+    },
+    {
+      "epoch": 1.67,
+      "learning_rate": 1.675619085055693e-05,
+      "loss": 1.9396,
+      "step": 1030
+    },
+    {
+      "epoch": 1.68,
+      "learning_rate": 1.6587098388746486e-05,
+      "loss": 1.9416,
+      "step": 1035
+    },
+    {
+      "epoch": 1.68,
+      "learning_rate": 1.6418256832286816e-05,
+      "loss": 1.9382,
+      "step": 1040
+    },
+    {
+      "epoch": 1.69,
+      "learning_rate": 1.6249678593856288e-05,
+      "loss": 1.9747,
+      "step": 1045
+    },
+    {
+      "epoch": 1.7,
+      "learning_rate": 1.6081376066774964e-05,
+      "loss": 1.8799,
+      "step": 1050
+    },
+    {
+      "epoch": 1.71,
+      "learning_rate": 1.591336162409352e-05,
+      "loss": 1.8957,
+      "step": 1055
+    },
+    {
+      "epoch": 1.72,
+      "learning_rate": 1.5745647617683627e-05,
+      "loss": 1.8921,
+      "step": 1060
+    },
+    {
+      "epoch": 1.73,
+      "learning_rate": 1.557824637732983e-05,
+      "loss": 1.9406,
+      "step": 1065
+    },
+    {
+      "epoch": 1.73,
+      "learning_rate": 1.5411170209823177e-05,
+      "loss": 1.9282,
+      "step": 1070
+    },
+    {
+      "epoch": 1.74,
+      "learning_rate": 1.5244431398056392e-05,
+      "loss": 1.8621,
+      "step": 1075
+    },
+    {
+      "epoch": 1.75,
+      "learning_rate": 1.5078042200120933e-05,
+      "loss": 1.9375,
+      "step": 1080
+    },
+    {
+      "epoch": 1.76,
+      "learning_rate": 1.4912014848405771e-05,
+      "loss": 1.8779,
+      "step": 1085
+    },
+    {
+      "epoch": 1.77,
+      "learning_rate": 1.4746361548698151e-05,
+      "loss": 1.9353,
+      "step": 1090
+    },
+    {
+      "epoch": 1.77,
+      "learning_rate": 1.4581094479286234e-05,
+      "loss": 1.9255,
+      "step": 1095
+    },
+    {
+      "epoch": 1.78,
+      "learning_rate": 1.4416225790063784e-05,
+      "loss": 1.9163,
+      "step": 1100
+    },
+    {
+      "epoch": 1.79,
+      "learning_rate": 1.4251767601636965e-05,
+      "loss": 1.9314,
+      "step": 1105
+    },
+    {
+      "epoch": 1.8,
+      "learning_rate": 1.4087732004433258e-05,
+      "loss": 1.8751,
+      "step": 1110
+    },
+    {
+      "epoch": 1.81,
+      "learning_rate": 1.3924131057812642e-05,
+      "loss": 1.8934,
+      "step": 1115
+    },
+    {
+      "epoch": 1.81,
+      "learning_rate": 1.376097678918098e-05,
+      "loss": 1.9148,
+      "step": 1120
+    },
+    {
+      "epoch": 1.82,
+      "learning_rate": 1.3598281193105858e-05,
+      "loss": 1.8754,
+      "step": 1125
+    },
+    {
+      "epoch": 1.83,
+      "learning_rate": 1.3436056230434747e-05,
+      "loss": 1.9183,
+      "step": 1130
+    },
+    {
+      "epoch": 1.84,
+      "learning_rate": 1.3274313827415678e-05,
+      "loss": 1.9236,
+      "step": 1135
+    },
+    {
+      "epoch": 1.85,
+      "learning_rate": 1.3113065874820506e-05,
+      "loss": 1.889,
+      "step": 1140
+    },
+    {
+      "epoch": 1.85,
+      "learning_rate": 1.295232422707068e-05,
+      "loss": 1.8898,
+      "step": 1145
+    },
+    {
+      "epoch": 1.86,
+      "learning_rate": 1.2792100701365794e-05,
+      "loss": 1.8991,
+      "step": 1150
+    },
+    {
+      "epoch": 1.87,
+      "learning_rate": 1.2632407076814794e-05,
+      "loss": 1.9559,
+      "step": 1155
+    },
+    {
+      "epoch": 1.88,
+      "learning_rate": 1.2473255093570039e-05,
+      "loss": 1.9048,
+      "step": 1160
+    },
+    {
+      "epoch": 1.89,
+      "learning_rate": 1.2314656451964196e-05,
+      "loss": 1.859,
+      "step": 1165
+    },
+    {
+      "epoch": 1.9,
+      "learning_rate": 1.2156622811650043e-05,
+      "loss": 1.8825,
+      "step": 1170
+    },
+    {
+      "epoch": 1.9,
+      "learning_rate": 1.1999165790743338e-05,
+      "loss": 1.9094,
+      "step": 1175
+    },
+    {
+      "epoch": 1.91,
+      "learning_rate": 1.1842296964968652e-05,
+      "loss": 1.937,
+      "step": 1180
+    },
+    {
+      "epoch": 1.92,
+      "learning_rate": 1.1686027866808394e-05,
+      "loss": 1.8838,
+      "step": 1185
+    },
+    {
+      "epoch": 1.93,
+      "learning_rate": 1.1530369984654936e-05,
+      "loss": 1.9023,
+      "step": 1190
+    },
+    {
+      "epoch": 1.94,
+      "learning_rate": 1.1375334761966074e-05,
+      "loss": 1.9099,
+      "step": 1195
+    },
+    {
+      "epoch": 1.94,
+      "learning_rate": 1.122093359642372e-05,
+      "loss": 1.9058,
+      "step": 1200
+    },
+    {
+      "epoch": 1.95,
+      "learning_rate": 1.1067177839095957e-05,
+      "loss": 1.9359,
+      "step": 1205
+    },
+    {
+      "epoch": 1.96,
+      "learning_rate": 1.0914078793602601e-05,
+      "loss": 1.8897,
+      "step": 1210
+    },
+    {
+      "epoch": 1.97,
+      "learning_rate": 1.0761647715284139e-05,
+      "loss": 1.9341,
+      "step": 1215
+    },
+    {
+      "epoch": 1.98,
+      "learning_rate": 1.0609895810374304e-05,
+      "loss": 1.876,
+      "step": 1220
+    },
+    {
+      "epoch": 1.98,
+      "learning_rate": 1.0458834235176225e-05,
+      "loss": 1.8287,
+      "step": 1225
+    },
+    {
+      "epoch": 1.99,
+      "learning_rate": 1.0308474095242267e-05,
+      "loss": 1.8523,
+      "step": 1230
+    },
+    {
+      "epoch": 2.0,
+      "learning_rate": 1.0128983382202781e-05,
+      "loss": 2.0887,
+      "step": 1235
+    },
+    {
+      "epoch": 2.01,
+      "learning_rate": 9.980205236069665e-06,
+      "loss": 1.4855,
+      "step": 1240
+    },
+    {
+      "epoch": 2.02,
+      "learning_rate": 9.832163712437392e-06,
+      "loss": 1.4915,
+      "step": 1245
+    },
+    {
+      "epoch": 2.03,
+      "learning_rate": 9.684869694834003e-06,
+      "loss": 1.4679,
+      "step": 1250
+    },
+    {
+      "epoch": 2.03,
+      "learning_rate": 9.538334011833363e-06,
+      "loss": 1.4298,
+      "step": 1255
+    },
+    {
+      "epoch": 2.04,
+      "learning_rate": 9.392567436259034e-06,
+      "loss": 1.4018,
+      "step": 1260
+    },
+    {
+      "epoch": 2.05,
+      "learning_rate": 9.247580684392345e-06,
+      "loss": 1.4642,
+      "step": 1265
+    },
+    {
+      "epoch": 2.06,
+      "learning_rate": 9.10338441518453e-06,
+      "loss": 1.4434,
+      "step": 1270
+    },
+    {
+      "epoch": 2.07,
+      "learning_rate": 8.959989229473125e-06,
+      "loss": 1.4574,
+      "step": 1275
+    },
+    {
+      "epoch": 2.07,
+      "learning_rate": 8.817405669202619e-06,
+      "loss": 1.4256,
+      "step": 1280
+    },
+    {
+      "epoch": 2.08,
+      "learning_rate": 8.675644216649478e-06,
+      "loss": 1.4539,
+      "step": 1285
+    },
+    {
+      "epoch": 2.09,
+      "learning_rate": 8.534715293651492e-06,
+      "loss": 1.5016,
+      "step": 1290
+    },
+    {
+      "epoch": 2.1,
+      "learning_rate": 8.39462926084159e-06,
+      "loss": 1.4738,
+      "step": 1295
+    },
+    {
+      "epoch": 2.11,
+      "learning_rate": 8.255396416886194e-06,
+      "loss": 1.4265,
+      "step": 1300
+    },
+    {
+      "epoch": 2.11,
+      "learning_rate": 8.117026997728079e-06,
+      "loss": 1.4235,
+      "step": 1305
+    },
+    {
+      "epoch": 2.12,
+      "learning_rate": 7.979531175833828e-06,
+      "loss": 1.5084,
+      "step": 1310
+    },
+    {
+      "epoch": 2.13,
+      "learning_rate": 7.842919059446046e-06,
+      "loss": 1.4426,
+      "step": 1315
+    },
+    {
+      "epoch": 2.14,
+      "learning_rate": 7.707200691840173e-06,
+      "loss": 1.4797,
+      "step": 1320
+    },
+    {
+      "epoch": 2.15,
+      "learning_rate": 7.572386050586196e-06,
+      "loss": 1.4309,
+      "step": 1325
+    },
+    {
+      "epoch": 2.16,
+      "learning_rate": 7.438485046815078e-06,
+      "loss": 1.4505,
+      "step": 1330
+    },
+    {
+      "epoch": 2.16,
+      "learning_rate": 7.305507524490145e-06,
+      "loss": 1.4734,
+      "step": 1335
+    },
+    {
+      "epoch": 2.17,
+      "learning_rate": 7.1734632596834106e-06,
+      "loss": 1.397,
+      "step": 1340
+    },
+    {
+      "epoch": 2.18,
+      "learning_rate": 7.042361959856825e-06,
+      "loss": 1.4341,
+      "step": 1345
+    },
+    {
+      "epoch": 2.19,
+      "learning_rate": 6.912213263148673e-06,
+      "loss": 1.4599,
+      "step": 1350
+    },
+    {
+      "epoch": 2.2,
+      "learning_rate": 6.783026737664942e-06,
+      "loss": 1.4466,
+      "step": 1355
+    },
+    {
+      "epoch": 2.2,
+      "learning_rate": 6.654811880775973e-06,
+      "loss": 1.4435,
+      "step": 1360
+    },
+    {
+      "epoch": 2.21,
+      "learning_rate": 6.527578118418187e-06,
+      "loss": 1.4597,
+      "step": 1365
+    },
+    {
+      "epoch": 2.22,
+      "learning_rate": 6.401334804401171e-06,
+      "loss": 1.4217,
+      "step": 1370
+    },
+    {
+      "epoch": 2.23,
+      "learning_rate": 6.276091219719984e-06,
+      "loss": 1.4477,
+      "step": 1375
+    },
+    {
+      "epoch": 2.24,
+      "learning_rate": 6.151856571872854e-06,
+      "loss": 1.4716,
+      "step": 1380
+    },
+    {
+      "epoch": 2.24,
+      "learning_rate": 6.028639994184277e-06,
+      "loss": 1.4398,
+      "step": 1385
+    },
+    {
+      "epoch": 2.25,
+      "learning_rate": 5.906450545133564e-06,
+      "loss": 1.4442,
+      "step": 1390
+    },
+    {
+      "epoch": 2.26,
+      "learning_rate": 5.785297207688905e-06,
+      "loss": 1.4506,
+      "step": 1395
+    },
+    {
+      "epoch": 2.27,
+      "learning_rate": 5.665188888646935e-06,
+      "loss": 1.4123,
+      "step": 1400
+    },
+    {
+      "epoch": 2.28,
+      "learning_rate": 5.546134417977984e-06,
+      "loss": 1.456,
+      "step": 1405
+    },
+    {
+      "epoch": 2.28,
+      "learning_rate": 5.428142548176876e-06,
+      "loss": 1.4274,
+      "step": 1410
+    },
+    {
+      "epoch": 2.29,
+      "learning_rate": 5.311221953619514e-06,
+      "loss": 1.4062,
+      "step": 1415
+    },
+    {
+      "epoch": 2.3,
+      "learning_rate": 5.195381229925156e-06,
+      "loss": 1.427,
+      "step": 1420
+    },
+    {
+      "epoch": 2.31,
+      "learning_rate": 5.080628893324475e-06,
+      "loss": 1.4783,
+      "step": 1425
+    },
+    {
+      "epoch": 2.32,
+      "learning_rate": 4.9669733800334955e-06,
+      "loss": 1.4356,
+      "step": 1430
+    },
+    {
+      "epoch": 2.33,
+      "learning_rate": 4.854423045633392e-06,
+      "loss": 1.4809,
+      "step": 1435
+    },
+    {
+      "epoch": 2.33,
+      "learning_rate": 4.742986164456196e-06,
+      "loss": 1.4079,
+      "step": 1440
+    },
+    {
+      "epoch": 2.34,
+      "learning_rate": 4.632670928976501e-06,
+      "loss": 1.4884,
+      "step": 1445
+    },
+    {
+      "epoch": 2.35,
+      "learning_rate": 4.523485449209195e-06,
+      "loss": 1.4499,
+      "step": 1450
+    },
+    {
+      "epoch": 2.36,
+      "learning_rate": 4.415437752113223e-06,
+      "loss": 1.4065,
+      "step": 1455
+    },
+    {
+      "epoch": 2.37,
+      "learning_rate": 4.308535781001457e-06,
+      "loss": 1.4888,
+      "step": 1460
+    },
+    {
+      "epoch": 2.37,
+      "learning_rate": 4.202787394956769e-06,
+      "loss": 1.4707,
+      "step": 1465
+    },
+    {
+      "epoch": 2.38,
+      "learning_rate": 4.0982003682542146e-06,
+      "loss": 1.4426,
+      "step": 1470
+    },
+    {
+      "epoch": 2.39,
+      "learning_rate": 3.994782389789535e-06,
+      "loss": 1.3991,
+      "step": 1475
+    },
+    {
+      "epoch": 2.4,
+      "learning_rate": 3.892541062513853e-06,
+      "loss": 1.4187,
+      "step": 1480
+    },
+    {
+      "epoch": 2.41,
+      "learning_rate": 3.7914839028747507e-06,
+      "loss": 1.4248,
+      "step": 1485
+    },
+    {
+      "epoch": 2.41,
+      "learning_rate": 3.691618340263701e-06,
+      "loss": 1.447,
+      "step": 1490
+    },
+    {
+      "epoch": 2.42,
+      "learning_rate": 3.5929517164698436e-06,
+      "loss": 1.4394,
+      "step": 1495
+    },
+    {
+      "epoch": 2.43,
+      "learning_rate": 3.495491285140282e-06,
+      "loss": 1.4359,
+      "step": 1500
+    },
+    {
+      "epoch": 2.44,
+      "learning_rate": 3.399244211246779e-06,
+      "loss": 1.4752,
+      "step": 1505
+    },
+    {
+      "epoch": 2.45,
+      "learning_rate": 3.304217570559052e-06,
+      "loss": 1.4508,
+      "step": 1510
+    },
+    {
+      "epoch": 2.45,
+      "learning_rate": 3.2104183491245466e-06,
+      "loss": 1.4718,
+      "step": 1515
+    },
+    {
+      "epoch": 2.46,
+      "learning_rate": 3.117853442754879e-06,
+      "loss": 1.4514,
+      "step": 1520
+    },
+    {
+      "epoch": 2.47,
+      "learning_rate": 3.026529656518864e-06,
+      "loss": 1.399,
+      "step": 1525
+    },
+    {
+      "epoch": 2.48,
+      "learning_rate": 2.936453704242215e-06,
+      "loss": 1.4136,
+      "step": 1530
+    },
+    {
+      "epoch": 2.49,
+      "learning_rate": 2.8476322080139862e-06,
+      "loss": 1.4474,
+      "step": 1535
+    },
+    {
+      "epoch": 2.5,
+      "learning_rate": 2.760071697699729e-06,
+      "loss": 1.4542,
+      "step": 1540
+    },
+    {
+      "epoch": 2.5,
+      "learning_rate": 2.673778610461448e-06,
+      "loss": 1.4176,
+      "step": 1545
+    },
+    {
+      "epoch": 2.51,
+      "learning_rate": 2.588759290284337e-06,
+      "loss": 1.4471,
+      "step": 1550
+    },
+    {
+      "epoch": 2.52,
+      "learning_rate": 2.505019987510426e-06,
+      "loss": 1.4217,
+      "step": 1555
+    },
+    {
+      "epoch": 2.53,
+      "learning_rate": 2.4225668583790474e-06,
+      "loss": 1.4194,
+      "step": 1560
+    },
+    {
+      "epoch": 2.54,
+      "learning_rate": 2.3414059645742504e-06,
+      "loss": 1.3959,
+      "step": 1565
+    },
+    {
+      "epoch": 2.54,
+      "learning_rate": 2.261543272779192e-06,
+      "loss": 1.4689,
+      "step": 1570
+    },
+    {
+      "epoch": 2.55,
+      "learning_rate": 2.1829846542374565e-06,
+      "loss": 1.4568,
+      "step": 1575
+    },
+    {
+      "epoch": 2.56,
+      "learning_rate": 2.105735884321436e-06,
+      "loss": 1.451,
+      "step": 1580
+    },
+    {
+      "epoch": 2.57,
+      "learning_rate": 2.029802642107734e-06,
+      "loss": 1.4418,
+      "step": 1585
+    },
+    {
+      "epoch": 2.58,
+      "learning_rate": 1.9551905099596813e-06,
+      "loss": 1.4619,
+      "step": 1590
+    },
+    {
+      "epoch": 2.58,
+      "learning_rate": 1.8819049731169059e-06,
+      "loss": 1.4182,
+      "step": 1595
+    },
+    {
+      "epoch": 2.59,
+      "learning_rate": 1.809951419292104e-06,
+      "loss": 1.4095,
+      "step": 1600
+    },
+    {
+      "epoch": 2.6,
+      "learning_rate": 1.7393351382749424e-06,
+      "loss": 1.4397,
+      "step": 1605
+    },
+    {
+      "epoch": 2.61,
+      "learning_rate": 1.6700613215431549e-06,
+      "loss": 1.4747,
+      "step": 1610
+    },
+    {
+      "epoch": 2.62,
+      "learning_rate": 1.6021350618809184e-06,
+      "loss": 1.4356,
+      "step": 1615
+    },
+    {
+      "epoch": 2.62,
+      "learning_rate": 1.5355613530044089e-06,
+      "loss": 1.4381,
+      "step": 1620
+    },
+    {
+      "epoch": 2.63,
+      "learning_rate": 1.470345089194709e-06,
+      "loss": 1.4444,
+      "step": 1625
+    },
+    {
+      "epoch": 2.64,
+      "learning_rate": 1.4064910649379803e-06,
+      "loss": 1.469,
+      "step": 1630
+    },
+    {
+      "epoch": 2.65,
+      "learning_rate": 1.3440039745729894e-06,
+      "loss": 1.4427,
+      "step": 1635
+    },
+    {
+      "epoch": 2.66,
+      "learning_rate": 1.2828884119460105e-06,
+      "loss": 1.3941,
+      "step": 1640
+    },
+    {
+      "epoch": 2.67,
+      "learning_rate": 1.2231488700730742e-06,
+      "loss": 1.4452,
+      "step": 1645
+    },
+    {
+      "epoch": 2.67,
+      "learning_rate": 1.1647897408096886e-06,
+      "loss": 1.4236,
+      "step": 1650
+    },
+    {
+      "epoch": 2.68,
+      "learning_rate": 1.107815314527929e-06,
+      "loss": 1.4538,
+      "step": 1655
+    },
+    {
+      "epoch": 2.69,
+      "learning_rate": 1.0522297798010594e-06,
+      "loss": 1.4112,
+      "step": 1660
+    },
+    {
+      "epoch": 2.7,
+      "learning_rate": 9.980372230955693e-07,
+      "loss": 1.4808,
+      "step": 1665
+    },
+    {
+      "epoch": 2.71,
+      "learning_rate": 9.452416284707743e-07,
+      "loss": 1.4509,
+      "step": 1670
+    },
+    {
+      "epoch": 2.71,
+      "learning_rate": 8.938468772859132e-07,
+      "loss": 1.4414,
+      "step": 1675
+    },
+    {
+      "epoch": 2.72,
+      "learning_rate": 8.438567479147975e-07,
+      "loss": 1.4203,
+      "step": 1680
+    },
+    {
+      "epoch": 2.73,
+      "learning_rate": 7.952749154680405e-07,
+      "loss": 1.4294,
+      "step": 1685
+    },
+    {
+      "epoch": 2.74,
+      "learning_rate": 7.481049515228811e-07,
+      "loss": 1.4136,
+      "step": 1690
+    },
+    {
+      "epoch": 2.75,
+      "learning_rate": 7.023503238606122e-07,
+      "loss": 1.4316,
+      "step": 1695
+    },
+    {
+      "epoch": 2.75,
+      "learning_rate": 6.580143962116281e-07,
+      "loss": 1.4645,
+      "step": 1700
+    },
+    {
+      "epoch": 2.76,
+      "learning_rate": 6.151004280081574e-07,
+      "loss": 1.4692,
+      "step": 1705
+    },
+    {
+      "epoch": 2.77,
+      "learning_rate": 5.736115741446146e-07,
+      "loss": 1.4408,
+      "step": 1710
+    },
+    {
+      "epoch": 2.78,
+      "learning_rate": 5.335508847456794e-07,
+      "loss": 1.4552,
+      "step": 1715
+    },
+    {
+      "epoch": 2.79,
+      "learning_rate": 4.949213049420576e-07,
+      "loss": 1.4657,
+      "step": 1720
+    },
+    {
+      "epoch": 2.79,
+      "learning_rate": 4.577256746539638e-07,
+      "loss": 1.4189,
+      "step": 1725
+    },
+    {
+      "epoch": 2.8,
+      "learning_rate": 4.2196672838233257e-07,
+      "loss": 1.4573,
+      "step": 1730
+    },
+    {
+      "epoch": 2.81,
+      "learning_rate": 3.876470950078037e-07,
+      "loss": 1.4382,
+      "step": 1735
+    },
+    {
+      "epoch": 2.82,
+      "learning_rate": 3.5476929759743927e-07,
+      "loss": 1.4272,
+      "step": 1740
+    },
+    {
+      "epoch": 2.83,
+      "learning_rate": 3.233357532192494e-07,
+      "loss": 1.4866,
+      "step": 1745
+    },
+    {
+      "epoch": 2.84,
+      "learning_rate": 2.933487727644813e-07,
+      "loss": 1.4132,
+      "step": 1750
+    },
+    {
+      "epoch": 2.84,
+      "learning_rate": 2.648105607777507e-07,
+      "loss": 1.4498,
+      "step": 1755
+    },
+    {
+      "epoch": 2.85,
+      "learning_rate": 2.3772321529494712e-07,
+      "loss": 1.4505,
+      "step": 1760
+    },
+    {
+      "epoch": 2.86,
+      "learning_rate": 2.1208872768901713e-07,
+      "loss": 1.4338,
+      "step": 1765
+    },
+    {
+      "epoch": 2.87,
+      "learning_rate": 1.8790898252354583e-07,
+      "loss": 1.4299,
+      "step": 1770
+    },
+    {
+      "epoch": 2.88,
+      "learning_rate": 1.6518575741421904e-07,
+      "loss": 1.4378,
+      "step": 1775
+    },
+    {
+      "epoch": 2.88,
+      "learning_rate": 1.4392072289814319e-07,
+      "loss": 1.4323,
+      "step": 1780
+    },
+    {
+      "epoch": 2.89,
+      "learning_rate": 1.241154423110169e-07,
+      "loss": 1.4144,
+      "step": 1785
+    },
+    {
+      "epoch": 2.9,
+      "learning_rate": 1.0577137167221863e-07,
+      "loss": 1.4343,
+      "step": 1790
+    },
+    {
+      "epoch": 2.91,
+      "learning_rate": 8.88898595777543e-08,
+      "loss": 1.4625,
+      "step": 1795
+    },
+    {
+      "epoch": 2.92,
+      "learning_rate": 7.347214710111239e-08,
+      "loss": 1.3614,
+      "step": 1800
+    },
+    {
+      "epoch": 2.92,
+      "learning_rate": 5.951936770202782e-08,
+      "loss": 1.4099,
+      "step": 1805
+    },
+    {
+      "epoch": 2.93,
+      "learning_rate": 4.7032547143155417e-08,
+      "loss": 1.4601,
+      "step": 1810
+    },
+    {
+      "epoch": 2.94,
+      "learning_rate": 3.60126034146524e-08,
+      "loss": 1.4231,
+      "step": 1815
+    },
+    {
+      "epoch": 2.95,
+      "learning_rate": 2.6460346666696835e-08,
+      "loss": 1.4549,
+      "step": 1820
+    },
+    {
+      "epoch": 2.96,
+      "learning_rate": 1.8376479149926353e-08,
+      "loss": 1.4122,
+      "step": 1825
+    },
+    {
+      "epoch": 2.96,
+      "learning_rate": 1.176159516380837e-08,
+      "loss": 1.3961,
+      "step": 1830
+    },
+    {
+      "epoch": 2.97,
+      "learning_rate": 6.616181012955025e-09,
+      "loss": 1.4652,
+      "step": 1835
+    },
+    {
+      "epoch": 2.98,
+      "learning_rate": 2.9406149713628874e-09,
+      "loss": 1.425,
+      "step": 1840
+    },
+    {
+      "epoch": 2.99,
+      "learning_rate": 7.351672546129785e-10,
+      "loss": 1.4916,
+      "step": 1845
+    },
+    {
+      "epoch": 3.0,
+      "learning_rate": 0.0,
+      "loss": 1.3883,
+      "step": 1850
+    },
+    {
+      "epoch": 3.0,
+      "step": 1851,
+      "total_flos": 1.8340911973547377e+18,
+      "train_loss": 2.0081952639620475,
+      "train_runtime": 43810.4873,
+      "train_samples_per_second": 5.415,
+      "train_steps_per_second": 0.042
+    }
+  ],
+  "max_steps": 1851,
+  "num_train_epochs": 3,
+  "total_flos": 1.8340911973547377e+18,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b8aabfc07bd4f8b579a13e3c5ed3e8dbbef17b9bcbe8603b010a6c6aac8df2d
+size 4207

zero_to_fp32.py

@@ -0,0 +1,484 @@
+#!/usr/bin/env python
+
+# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+import deepspeed
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION,
+                                            OPTIMIZER_STATE_DICT,
+                                            PARAM_SHAPES,
+                                            SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS,
+                                            ZERO_STAGE,
+                                            PARTITION_COUNT,
+                                            PARAM_SHAPES,
+                                            BUFFER_NAMES)
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage == 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_optim_files(checkpoint_dir):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
+                                                "*_optim_states.pt")),
+                         key=natural_keys)
+
+    if len(optim_files) == 0:
+        raise FileNotFoundError(
+            f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
+
+    return optim_files
+
+
+def parse_model_state(file):
+    state_dict = torch.load(file, map_location=device)
+
+    if BUFFER_NAMES not in state_dict:
+        raise ValueError(f"{file} is not a model state checkpoint")
+    buffer_names = state_dict[BUFFER_NAMES]
+    if debug:
+        print("Found buffers:", buffer_names)
+
+    # recover just the buffers while restoring them to fp32 if they were saved in fp16
+    buffers = {
+        k: v.float()
+        for k,
+        v in state_dict["module"].items() if k in buffer_names
+    }
+    param_shapes = state_dict[PARAM_SHAPES]
+
+    ds_version = state_dict.get(DS_VERSION, None)
+
+    return buffers, param_shapes, ds_version
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dicts.append(torch.load(f, map_location=device))
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage == 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage == 2:
+        fp32_flat_groups = [
+            state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
+            for i in range(len(state_dicts))
+        ]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key],
+                      0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(
+        f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
+    buffers, param_shapes, ds_version = parse_model_state(model_file)
+    print(f'Parsing checkpoint created by deepspeed=={ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(
+                    f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum([
+        full_single_fp32_vector.numel()
+        for full_single_fp32_vector in merged_single_partition_of_fp32_groups
+    ])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum(
+            [sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel()
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(
+                    f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
+                )
+            state_dict[name] = full_single_fp32_vector.narrow(
+                0,
+                offset,
+                unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(
+                f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(
+        f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0,
+                                             offset,
+                                             partitioned_numel)
+                  for i in range(world_size)),
+            0).narrow(0,
+                      0,
+                      unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(
+            f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(
+        f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "checkpoint_dir",
+        type=str,
+        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help=
+        "path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
+    )
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)